AI Content Recommendations Implementation

Implementing AI-Powered Content Recommendations

Content recommendations keep users engaged and increase page depth. The approach depends on available behavioral data: for a new site without history — content-based filtering with embeddings; for a site with thousands of users and events — collaborative filtering or hybrid systems.

Choosing an Approach

Content-Based: Similar Content via Embeddings

Fastest approach — similar articles based on vector distance:

import OpenAI from 'openai';
import { sql } from '@vercel/postgres';

const openai = new OpenAI();

// Index on article publish
async function indexArticle(article) {
  const textToEmbed = [
    article.title,
    article.excerpt,
    article.tags.join(', '),
    article.body.slice(0, 2000), // first 2000 chars
  ].join('\n\n');

  const { data: [{ embedding }] } = await openai.embeddings.create({
    model: 'text-embedding-3-small',
    input: textToEmbed,
  });

  await sql`
    UPDATE articles
    SET embedding = ${JSON.stringify(embedding)}::vector
    WHERE id = ${article.id}
  `;
}

// Get similar articles
async function getSimilarArticles(articleId, limit = 6) {
  const result = await sql`
    WITH source AS (
      SELECT embedding FROM articles WHERE id = ${articleId}
    )
    SELECT
      a.id, a.title, a.slug, a.excerpt, a.published_at,
      a.category, a.read_time,
      1 - (a.embedding <=> source.embedding) AS similarity
    FROM articles a, source
    WHERE a.id != ${articleId}
      AND a.published = true
      AND a.embedding IS NOT NULL
    ORDER BY a.embedding <=> source.embedding
    LIMIT ${limit}
  `;

  return result.rows;
}

Collaborative Filtering: "Users Like You Read"

Matrix factorization via implicit feedback (views, time on page):

# Python script for periodic training (cron)
import implicit
import numpy as np
from scipy.sparse import csr_matrix
import pickle

def train_collaborative_model():
    # Load events: user_id, article_id, weight
    # weight = 1 (view) + 2 (scroll 50%) + 5 (read to end) + 10 (shared)
    events = fetch_events_from_db()

    users = {u: i for i, u in enumerate(events['user_id'].unique())}
    items = {a: i for i, a in enumerate(events['article_id'].unique())}

    rows = events['user_id'].map(users)
    cols = events['article_id'].map(items)
    data = events['weight']

    matrix = csr_matrix((data, (rows, cols)))

    model = implicit.als.AlternatingLeastSquares(
        factors=128,
        regularization=0.01,
        iterations=50,
        use_gpu=False,
    )
    model.fit(matrix)

    # Save model and mappings
    with open('/models/collab_model.pkl', 'wb') as f:
        pickle.dump({ 'model': model, 'users': users, 'items': items }, f)

// Node.js: get recommendations via Python service
async function getCollaborativeRecs(userId, limit = 10) {
  const response = await fetch('http://ml-service:5000/recommend', {
    method: 'POST',
    body: JSON.stringify({ user_id: userId, limit }),
  });
  return response.json();
}

Hybrid System with Personalization

Combine content-based and collaborative signals:

async function getPersonalizedRecommendations(userId, currentArticleId) {
  const [contentBased, collaborative, trending] = await Promise.all([
    getSimilarArticles(currentArticleId, 10),
    getCollaborativeRecs(userId, 10),
    getTrendingArticles(10), // by views in last 24h
  ]);

  // Merge with weights
  const scores = new Map();

  contentBased.forEach((article, i) => {
    scores.set(article.id, (scores.get(article.id) || 0) + (10 - i) * 0.4);
  });

  collaborative.forEach((article, i) => {
    scores.set(article.id, (scores.get(article.id) || 0) + (10 - i) * 0.5);
  });

  trending.forEach((article, i) => {
    scores.set(article.id, (scores.get(article.id) || 0) + (10 - i) * 0.1);
  });

  // Sort by total score
  const allArticleIds = [...scores.keys()];
  const articles = await fetchArticlesByIds(allArticleIds);

  return articles
    .map(a => ({ ...a, score: scores.get(a.id) }))
    .sort((a, b) => b.score - a.score)
    .slice(0, 6);
}

LLM-Based Recommendations with Explanation

For smarter matching and personalized explanations:

async function getLLMRecommendations(user, readHistory, availableArticles) {
  const userProfile = `
    Read: ${readHistory.map(a => a.title).join(', ')}
    Interest categories: ${getTopCategories(readHistory).join(', ')}
    Average reading time: ${user.avgReadTime} min
  `;

  const articlesList = availableArticles.slice(0, 20).map(a =>
    `ID:${a.id} | ${a.title} | ${a.category} | ${a.tags.join(',')}`
  ).join('\n');

  const response = await openai.chat.completions.create({
    model: 'gpt-4o-mini',
    response_format: { type: 'json_object' },
    messages: [
      {
        role: 'system',
        content: 'You are a recommendation system. Respond in JSON: { recommendations: [{id, reason}] }',
      },
      {
        role: 'user',
        content: `Profile: ${userProfile}\n\nAvailable articles:\n${articlesList}\n\nSelect 4 most relevant for this user.`,
      },
    ],
    max_tokens: 400,
  });

  const { recommendations } = JSON.parse(response.choices[0].message.content);

  // Enrich with DB data
  return Promise.all(recommendations.map(async rec => ({
    ...await fetchArticle(rec.id),
    reason: rec.reason, // "You've read similar material about React"
  })));
}

Event Tracking

Behavioral data is the foundation for improving recommendations:

// Client tracker
class ReadingTracker {
  constructor(articleId) {
    this.articleId = articleId;
    this.startTime = Date.now();
    this.maxScroll = 0;
    this.trackScroll();
  }

  trackScroll() {
    const observer = new IntersectionObserver((entries) => {
      entries.forEach(entry => {
        if (entry.isIntersecting) {
          const progress = entry.target.dataset.progress;
          if (progress > this.maxScroll) {
            this.maxScroll = progress;
            this.sendEvent('scroll', { progress });
          }
        }
      });
    });

    document.querySelectorAll('[data-progress]').forEach(el => observer.observe(el));
  }

  async sendEvent(type, data = {}) {
    navigator.sendBeacon('/api/track', JSON.stringify({
      type,
      articleId: this.articleId,
      timeOnPage: Date.now() - this.startTime,
      ...data,
    }));
  }
}

Caching Recommendations

Recommendations are expensive, so cache them:

async function getCachedRecommendations(userId, articleId) {
  const cacheKey = `recs:${userId}:${articleId}`;
  const cached = await redis.get(cacheKey);

  if (cached) return JSON.parse(cached);

  const recs = await getPersonalizedRecommendations(userId, articleId);
  await redis.setex(cacheKey, 3600, JSON.stringify(recs)); // 1 hour

  return recs;
}

Timeline

• Content-based recommendations via pgvector — 3–4 days
• Event tracking + behavior analytics — plus 2 days
• Collaborative filtering (implicit ALS) — plus 3–4 days
• Hybrid system with LLM explanations — 2–3 weeks full cycle
• A/B testing algorithms — plus 2–3 days